Comparative Machine Learning Modeling of Composite Shields for Secondary Radiation Dose Prediction in Passive Proton Therapy

Passive proton therapy generates secondary neutrons and photons that require effective shielding to ensure radiation protection. Although Monte Carlo simulations provide accurate dose calculations, they are time-consuming for shielding optimization studies. In this work, Artificial Neural Network (ANN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) models were developed and compared for rapid prediction of the total equivalent dose behind composite shields. A dataset consisting of 432 shielding configurations was generated using MCNPX simulations for a 250 MeV proton beam. Shielding thickness, degrader thickness, and the weight fractions of steel-scrap and serpentine in concrete were considered as input parameters, while the total equivalent dose of secondary radiation was used as the output parameter. All three machine-learning models achieved excellent predictive performance, with R² values close to 0.99 for both training and testing datasets. While RF and XGBoost showed slightly lower prediction errors within the available dataset, independent validation using previously unseen shielding configurations demonstrated that ANN provided more accurate predictions and better generalization capability. Feature importance analysis indicated that shielding thickness is the dominant parameter affecting equivalent dose, followed by degrader thickness.